The Pulse

ElevenLabs releases Scribe v3 transcription models, citing fewer errors

ElevenLabs has released Scribe v3 for recorded audio and Scribe v3 Live for real-time transcription. The company says its tests found lower error rates than their predecessors across 102 languages.

AI.info Team · · fact-checked · against 1 source ↓ · Published 11 Nov 2026

Scribe v3 produces 27% fewer word errors than its predecessor across 102 languages, according to ElevenLabs, which released the recorded-audio transcription model alongside a real-time version on November 10. The company says Scribe v3 Live also reduced errors compared with Scribe v2 Realtime and adds speaker labeling and improved turn detection.

Both models are available through ElevenAPI and are the defaults in the ElevenLabs app. Scribe v3 Live is also available in ElevenAgents, including a version optimized for medical conversations.

How the models were tested

ElevenLabs measured multilingual accuracy using an internal version of the FLEURS benchmark, which consists of short, clean read speech in 102 languages. Because those clips each contain a single speaker, the company built a separate dataset of real meetings, agent calls and voice notes across 16 languages to evaluate speaker labeling.

For the live model, semantic word error fell by 43% and standard word error by 31% across 102 languages compared with Scribe v2 Realtime, according to the company. Semantic word error counts an error when it changes the meaning of a transcript, rather than counting differences in formatting that do not change meaning.

ElevenLabs says it scored each language separately before averaging, giving each of the 102 languages equal weight. It cautions that results measured by pooling clips into a single score are not directly comparable. The company plans to open-source the real-world dataset.

Speaker labels and conversation turns

The recorded-audio model made 12% fewer speaker-labeling errors overall than Scribe v2 in the company’s tests, and 17% fewer on recordings with three or more speakers. For the live model, ElevenLabs says it added speaker diarization, which identifies who said what, for the first time.

On the English portion of that dataset, Scribe v3 Live had a 3.1% diarization error rate, compared with 5.9% to 7.3% for the other real-time models tested, the company reports. The live scores were calculated after aligning each model’s transcript to a reference to isolate speaker-labeling errors; ElevenLabs says they cannot be compared with its recorded-audio diarization results.

Scribe v3 Live also uses conversation context and audio to detect when a speaker has finished a turn. The feature, already used in ElevenAgents, is now available in ElevenAPI.

Controls and deployment

Scribe v3 can infer possible languages in a recording and lets users choose whether numbers appear as digits or words. A keyterm prompt can hold up to 1,000 terms, while transcript edits can apply formatting instructions written in plain English.

The live model adds a multilingual mode for conversations that switch languages mid-sentence. It also offers a keyterm library containing thousands of possible terms, which the model consults when it hears a likely match.

Both models run in the US by default, with isolated environments in the EU, India and Singapore offered to enterprise customers with data residency requirements. They are available for private deployment and are HIPAA-eligible with a Business Associate Agreement. Zero Retention Mode deletes audio and text when a request completes.